Ontology-grounded AI is the right idea — but today's ontologies inform the model and hope. We ran a deterministic enforcer over two real ontologies, including OMOP, the clinical-research standard: 39 tables and 94 relationships of rulebook from the published spec, zero authoring.
We measured fan-out queries on real benchmark databases: results inflated up to 272×, and the wrong query did up to 8,000× more row-work than the right one — billed to your warehouse. What a deterministic gate buys a text-to-SQL (NL2SQL) product, in dollars.
The industry's default safety pattern — LLM-as-judge subagents — spends tokens on every query, every retry, every agent. As pricing converges toward compute cost, deterministic checks are the only guardrails with zero exposure. Spend tokens on ambiguity, not arithmetic.
We ran a deterministic semantic check over 2,568 gold queries from Spider and BIRD: zero false alarms, a schema defect filed upstream, and one gold answer proven wrong by 8× — by executing the benchmark's own database.